arXiv subjects

Adrian Lozano-Duran

Publications and source records attributed to Adrian Lozano-Duran.

14 recordsLinked to original sources

arXiv2026

Ladder of information limits on prediction for reduced-order models

What can and cannot be predicted by a model when only limited information is available? We answer this question by constructing a ladder of information limits that organizes prediction tasks according to their information requirements: from memoryless and memory-augmented trajectory forecasts to event prediction, stationary statistics, and generative laws. The analysis, which is independent of the model structure or architecture, identifies the minimum attainable error along with the information limitations that give rise to it. The approach distinguishes information hidden in unresolved variables from information recovered through time-delayed observations and connects these contributions to Mori--Zwanzig memory. The ladder also reveals why Lyapunov growth alone cannot characterize reduced-order prediction error, why stationary statistics may remain predictable even when individual trajectories become unpredictable, and why stochasticity can represent (but cannot recover) missing information. Numerical studies of the Kuramoto--Sivashinsky equation and the Lorenz system illustrate these results across the ladder. This unified perspective provides a principled language for understanding the fundamental limitations of ROMs for a given prediction task and for clarifying whether improved predictions require richer observations, greater precision in the input, or additional memory.

nlin.CD↗

arXiv2026

AeroJEPA: Learning Semantic Latent Representations for Scalable 3D Aerodynamic Field Modeling

Aerodynamic surrogate models are increasingly used to replace repeated high-fidelity CFD evaluations in many-query design settings, but current approaches still face two important limitations: they often scale poorly to the very large fields arising in realistic 3D aerodynamics, and they rarely produce latent representations that are directly useful for analysis and design. We introduce AeroJEPA, a Joint-Embedding Predictive Architecture for aerodynamic field modeling that addresses both issues. Rather than predicting the full flow field directly from geometry, AeroJEPA predicts a target latent representation of the flow from a context latent representation of the geometry and operating conditions, and optionally reconstructs the field through a continuous implicit decoder. This formulation decouples latent prediction from field resolution while encouraging the latent space to organize semantically. We evaluate AeroJEPA on two complementary datasets: HiLiftAeroML, which stresses the method in a high-fidelity regime with extremely large boundary-layer fields, and SuperWing, which tests large-scale generalization and latent-space optimization over a broad family of transonic wings. Across these benchmarks, AeroJEPA is competitive as a continuous surrogate for aerodynamic fields, scales naturally to high-resolution outputs, and learns context and predicted latents that encode geometry and aerodynamic quantities not used directly as supervision. We further show that the resulting latent space supports controlled interpolation, linear probing, concept-vector arithmetic, and a constrained design latent-optimization experiment. These results suggest that predictive latent learning is a promising direction for scalable and design-meaningful aerodynamic surrogate modeling.

cs.LG↗

arXiv2026

Unified scaling laws for turbulent boundary layers across flow regimes

We discover unified scaling laws for the mean wall shear stress and the mean velocity profile in turbulent boundary layers subject to favorable and adverse mean pressure gradients-including flows with separation and reattachment. We use the information-theoretic irreducible error theorem to identify, among all dimensionally consistent combinations, the dimensionless groups with maximal predictive power, without assuming any functional form. Two dimensionless variables suffice to describe the mean wall shear stress, while three characterize the mean velocity profile. The scaling laws depend exclusively on variables defined at a fixed streamwise location, demonstrating that judiciously chosen combinations of local quantities implicitly encode upstream history without requiring global parameters. The results are validated against a rich collection of cases and are shown to collapse mean quantities across flow regimes previously thought to require distinct treatments.

physics.flu-dyn↗

arXiv2026

Agentic Exploration of PDE Spaces using Latent Foundation Models for Parameterized Simulations

Flow physics and more broadly physical phenomena governed by partial differential equations (PDEs), are inherently continuous, high-dimensional and often chaotic in nature. Traditionally, researchers have explored these rich spatiotemporal PDE solution spaces using laboratory experiments and/or computationally expensive numerical simulations. This severely limits automated and large-scale exploration, unlike domains such as drug discovery or materials science, where discrete, tokenizable representations naturally interface with large language models. We address this by coupling multi-agent LLMs with latent foundation models (LFMs), a generative model over parametrised simulations, that learns explicit, compact and disentangled latent representations of flow fields, enabling continuous exploration across governing PDE parameters and boundary conditions. The LFM serves as an on-demand surrogate simulator, allowing agents to query arbitrary parameter configurations at negligible cost. A hierarchical agent architecture orchestrates exploration through a closed loop of hypothesis, experimentation, analysis and verification, with a tool-modular interface requiring no user support. Applied to flow past tandem cylinders at Re = 500, the framework autonomously evaluates over 1,600 parameter-location pairs and discovers divergent scaling laws: a regime-dependent two-mode structure for minimum displacement thickness and a robust linear scaling for maximum momentum thickness, with both landscapes exhibiting a dual-extrema structure that emerges at the near-wake to co-shedding regime transition. The coupling of the learned physical representations with agentic reasoning establishes a general paradigm for automated scientific discovery in PDE-governed systems.

cs.AI↗

arXiv2026

Machine-learning wall model of large-eddy simulation for low- and high-speed flows over rough surfaces

We present a wall model for large-eddy simulation that incorporates surface-roughness effects and is applicable across low- and high-speed flows, for both transitional and fully rough conditions. The model, implemented using an artificial neural network, is trained on a direct numerical simulation database of compressible turbulent channel flows over rough walls. The dataset contains 372 cases spanning a wide range of irregular roughness topographies, including Gaussian and Weibull distributions, Mach numbers 0~3.3, and friction Reynolds numbers 180~2000. We employ an information-theoretic, dimensionless learning method to identify the inputs with the highest predictive power for the dimensionless wall friction and wall heat flux. Predictions are accompanied by a confidence score derived from a spectrally normalized neural Gaussian process, which quantifies uncertainty in regions that deviate from the training dataset. The model performance is first evaluated a-priori on 110 turbulent channel flow cases, yielding prediction errors below 4%. The model is assessed a-posteriori in wall-modeled large-eddy simulations across diverse test cases. These include over 160 subsonic and supersonic turbulent channel flows with rough walls, a transonic high-pressure turbine (HPT) blade with Gaussian roughness, a high-speed compression ramp with sandpaper roughness, and three hypersonic blunt bodies with sand-grain roughness. Results show that the proposed wall model typically achieves a-posteriori predictive accuracy within 10% for wall shear stress and within 15% for wall heat flux, with high confidence in the channel flows and HPT blade cases. In the rough-wall compression ramp and hypersonic blunt bodies, the model captures the heating augmentation with errors ranging 0%~20%. In the cases with the highest errors, the reduced performance is correctly detected by a drop in the confidence score.

physics.flu-dyn↗

arXiv2026

X-CAL: Explaining latent causality in physical space for fluid mechanics

We present X-CAL, a pipeline that combines a $β$-variational autoencoder ($β$-VAE) with the synergistic-unique-redundant decomposition (SURD)~\cite{surd} approach for causality analysis to interpret low-dimensional latent representations of turbulent fluid flows. Combining $β$-VAE compression with SURD and SHAP (SHapley Additive exPlanations) yields interpretable latent representations and structure-level attributions in physical space, offering a general methodology for causal analysis of high-dimensional flows. Using direct numerical simulation (DNS) data of the flow around a wall-mounted square cylinder at $Re_h=2000$, we (i) learn a compact latent space with near-orthogonal variables, (ii) quantify directed information flows among these variables via the SURD approach, and (iii) map latent-space causality back to physical space through gradient-SHAP fields . By means of percolation analysis of the SHAP fields, we extract the coherent, time-resolved structures that most influence each latent variable. The analysis connects coherent structures with latent variables which are in turn associated with wake-boundary-layer interactions. This method enables translating the insight obtained through causal analysis in the latent space into interpretable phenomena in physical space.

physics.flu-dyn↗

arXiv2026

Data-Driven Reduced-Complexity Modeling of Fluid Flows: A Community Challenge

We introduce a community challenge designed to facilitate direct comparisons between data-driven methods for compression, forecasting, and sensing of complex aerospace flows. The challenge is organized into three tracks that target these complementary capabilities: compression (compact representations for large datasets), forecasting (predicting future flow states from a finite history), and sensing (inferring unmeasured flow states from limited measurements). Across these tracks, multiple challenges span diverse flow datasets and use cases, each emphasizing different model requirements. The challenge is open to anyone, and we invite broad participation to build a comprehensive and balanced picture of what works and where current methods fall short. To support fair comparisons, we provide standardized success metrics, evaluation tools, and baseline implementations, with one classical and one machine-learning baseline per challenge. Final assessments use blind tests on withheld data. We explicitly encourage negative results and careful analyses of limitations. Outcomes will be disseminated through an AIAA Journal Virtual Collection and invited presentations at AIAA conferences.

cs.LG↗

arXiv2025

General-purpose Data-driven Wall Model for Low-speed Flows Part I: Baseline Model

We present a general-purpose wall model for large-eddy simulation. The model builds on the building-block flow principle, leveraging essential physics from simple flows to train a generalizable model applicable across complex geometries and flow conditions. The model addresses key limitations of traditional equilibrium wall models (EQWM) and improves upon shortcomings of earlier building-block-based approaches. The model comprises four components: (i) a baseline wall model, (ii) an error model, (iii) a classifier, and (iv) a confidence score. The baseline model predicts the wall-shear stress, while the error model estimates epistemic errors and aleatoric errors, both used for uncertainty quantification. In Part I of this work, we present the baseline model, while the remaining three components are introduced in Part II. The baseline model is designed to capture a broad range of flow phenomena, including turbulence over curved walls and zero, adverse, and favorable mean pressure gradients, as well as flow separation and laminar flow. The problem is formulated as a regression task to predict wall shear stress using a neural network. Model inputs are localized in space and dimensionless, with their selection guided by information-theoretic criteria. Training data include, among other cases, a newly generated direct numerical simulation dataset of turbulent boundary layers under favorable and adverse PG conditions. Validation is carried out through both a priori and a posteriori tests. The a priori evaluation spans 140 diverse high-fidelity numerical datasets and experiments (67 training cases included), covering turbulent boundary layers, airfoils, Gaussian bumps, and full aircraft geometries, among others. We demonstrate that the baseline wall model outperforms the EQWM in 90% of test scenarios, while maintaining errors below 20% for 98% of the cases.

physics.flu-dyn↗

arXiv2025

The coherent structure of the energy cascade in isotropic turbulence

The energy cascade, i.e. the transfer of kinetic energy from large-scale to small-scale flow motions, has been the cornerstone of turbulence theories and models since the 1940s. However, understanding the spatial organization of the energy transfer has remained elusive. In this work, we answer the question: What are the characteristic flow patterns surrounding regions of intense energy transfer? To that end, we utilize numerical data of isotropic turbulence to investigate the three-dimensional spatial structure of the energy cascade in the inertial range. Our findings indicate that forward energy-transfer events are predominantly confined in the high strain-rate region created between two distinct zones of elevated enstrophy. On average, these zones manifest in the form of two hairpin-like shapes with opposing orientations. The mean velocity field associated with the energy transfer exhibits a saddle point topology when observed in the frame of reference local to the event. The analysis also shows that the primary driving mechanism for the cascade involves strain-rate self-amplification, which is responsible for 85% of the energy transfer, whereas vortex stretching accounts for less than 15%.

physics.flu-dyn↗

arXiv2024

Building-block flow model for computational fluids

We introduce a closure model for wall-modeled large-eddy simulation (WMLES), referred to as the Building-block Flow Model (BFM). The foundation of the model rests on the premise that a finite collection of simple flows encapsulates the essential physics necessary to predict more complex scenarios. The BFM is implemented using artificial neural networks and introduces five advancements within the framework of WMLES: (1) It is designed to predict multiple flow regimes (wall turbulence under zero, favorable, adverse mean-pressure-gradient, and separation); (2) It unifies the closure model at solid boundaries (i.e., the wall model) and the rest of the flow (i.e., the subgrid-scale model) into a single entity; (3) It ensures consistency with numerical schemes and gridding strategy by accounting for numerical errors; (4) It is directly applicable to arbitrary complex geometries; (5) It can be scaled up to model additional flow physics in the future if needed (e.g., shockwaves and laminar-to-turbulent transition). The BFM is utilized to predict key quantities of interest in turbulent channel and pipe flows, a Gaussian bump, a simplified aircraft, and a realistic aircraft in landing configuration. In all cases, the BFM demonstrates similar or superior capabilities in terms of accuracy and computational efficiency compared to previous state-of-the-art closure models.

physics.flu-dyn↗

arXiv2024

Machine-learning wall-model large-eddy simulation accounting for isotropic roughness under local equilibrium

We introduce a wall model (WM) for large-eddy simulation (LES) applicable to rough surfaces with Gaussian and non-Gaussian distributions for both transitionally and fully rough regimes. The model is applicable to arbitrary complex geometries where roughness elements are assumed to be underresolved. The wall model is implemented using a feedforward neural network, with the geometric properties of the roughness topology and near-wall flow quantities serving as input. The optimal set of non-dimensional input features is identified using information theory, selecting variables that maximize information about the output while minimizing redundancy among inputs. The model incorporates a confidence score based on Gaussian process modeling, enabling the detection of low model performance for unseen rough surfaces. The model is trained using a direct numerical simulation roughness database comprising approximately 200 cases. The roughness geometries for the database are selected from a large repository through active learning. This approach ensures that the rough surfaces incorporated into the database are the most informative. The model performance is evaluated both a-priori and a-posteriori in WMLES of turbulent channel flows with rough walls. Over 120 channel flow cases are considered, including untrained roughness geometries, roughness Reynolds numbers, and grid resolutions for both transitionally- and fully-rough regimes. The results show that the rough-wall model predicts the wall shear stress within 15% accuracy. The model is also assessed on a high-pressure turbine blade with two different rough surfaces. The WM predicts the skin friction and the mean velocity deficit within 10% accuracy except the region with shock waves.

physics.flu-dyn↗

arXiv2022

Numerical and modeling error assessment of large-eddy simulation using direct-numerical-simulation-aided large-eddy simulation

We study the numerical errors of large-eddy simulation (LES) in isotropic and wall-bounded turbulence. A direct-numerical-simulation (DNS)-aided LES formulation, where the subgrid-scale (SGS) term of the LES is computed by using filtered DNS data is introduced. We first verify that this formulation has zero error in the absence of commutation error between the filter and the differentiation operator of the numerical algorithm. This method allows the evaluation of the time evolution of numerical errors for various numerical schemes at grid resolutions relevant to LES. The analysis shows that the numerical errors are of the same order of magnitude as the modeling errors and often cancel each other. This supports the idea that supervised machine learning algorithms trained on filtered DNS data might not be suitable for robust SGS model development, as this approach disregards the existence of numerical errors in the system that accumulates over time. The assessment of errors in turbulent channel flow also identifies that numerical errors close to the wall dominate, which has implications for the development of wall models.

physics.flu-dyn↗

arXiv2020

Nonlinear mechanism of the self-sustaining process in the buffer and logarithmic layer of wall-bounded flows

The nonlinear mechanism in the self-sustaining process (SSP) of wall-bounded turbulence is investigated. Resolvent analysis is used to identify the principal forcing mode which produces the maximum amplification of the velocities in numerical simulations of the minimal channel for the buffer layer and a modified logarithmic (log) layer. The wavenumbers targeted in this study are those of the fundamental mode that is infinitely long in the streamwise direction and once periodic in the spanwise direction. The identified mode is then projected out from the nonlinear term of the Navier-Stokes equations at each time step from the simulation of the corresponding minimal channel. The results show that the removal of the principal forcing mode of the fundamental wavenumber can inhibit turbulence in both the buffer and log layer, with the effect being greater in the buffer layer. Removing other modes instead of the principal mode of the fundamental wavenumber only marginally affects the flow. Closer inspection of the dyadic interactions in the nonlinear term shows that contributions toward the principal forcing mode come from a limited set of wavenumber interactions. Using conditional averaging, the flow structures that are responsible for generating the nonlinear interaction to self-sustain turbulence are identified as spanwise rolls interacting with oblique streaks. This method, based on the equations of motion, validates the similarities in the SSP of the buffer and log layer, and characterises the underlying quadratic interactions in the SSP of the minimal channel.

physics.flu-dyn↗

arXiv2019

Resolvent-based estimation of space-time flow statistics

We develop a method to estimate space-time flow statistics from a limited set of known data. While previous work has focused on modeling spatial or temporal statistics independently, space-time statistics carry fundamental information about the physics and coherent motions of the flow and provide a starting point for low-order modeling and flow control efforts. The method is derived using a statistical interpretation of resolvent analysis. The central idea of our approach is to use known data to infer the statistics of the nonlinear terms that constitute a forcing on the linearized Navier-Stokes equations, which in turn imply values for the remaining unknown flow statistics through application of the resolvent operator. Rather than making an a priori rank-1 assumption, our method allows the known input data to select the most relevant portions of the resolvent operator for describing the data, making it well-suited for high-rank turbulent flows. We demonstrate the predictive capabilities of the method using two examples: the Ginzburg-Landau equation, which serves as a convenient model for a convectively unstable flow, and a turbulent channel flow at low Reynolds number.

physics.flu-dyn↗